Cohere AI has launched Rerank 5, the latest iteration of its retrieval model suite designed to reorder search results before passing data to large language models. Available in Pro and Fast tiers, Rerank 5 supports over 100 languages, a 32K context window, and semi-structured data formats like JSON, tables, and YAML. The models are available via Cohere’s API, Microsoft Foundry, Amazon SageMaker, and native framework integrations including LangChain and Elasticsearch.
A key feature in Rerank 5 is Context Focus, an optional setting that identifies and extracts specific relevant character spans within documents. Cohere reports that using Context Focus reduced token consumption per task by up to 41% while boosting accuracy from 39% to 52% across deep-search benchmarks. Additionally, Rerank 5 introduces calibrated relevance scores that output precise probability estimates, allowing systems to set explicit relevance thresholds rather than relying on fixed top-k document cutoffs.
According to Cohere’s internal benchmarks, Rerank 5 Pro achieved top scores across multiple evaluations, including 92.2 on ViDoRe V3 and 90.4 on BEIR. API pricing is set at $0.05 per 1 million tokens for the Pro tier and $0.02 per 1 million tokens for the Fast tier.
Why it matters
Reduces LLM inference costs and context window bloat for enterprise RAG applications by extracting exact relevant passage spans.
Improves agentic decision-making by providing calibrated probability scores, enabling reliable threshold filtering instead of arbitrary top-k limits.
Allows seamless drop-in upgrades for existing search pipelines without requiring expensive index re-building or data re-vectorization.
Source: marktechpost.com



